Economic Forecasting With Autoregressive Methods and Neural Networks

Abstract

Neural networks can forecast economic data with accuracy matching that of conventional autoregressive methods such as SARIMA and VAR. This study uses dense, recurrent, convolutional, and convnet/RNN hybrids to conduct time-series analysis of interest rates, consumer and producer prices, and labor market data. Training on 14 years of data, neural networks produce accurate 50-year forecasts. Gaps in these forecasts may reveal macroeconomic regime changes. Failures in otherwise accurate neural network forecasts may thus inform theoretical economic hypotheses through unsupervised machine learning.